Gradient-guided Attention Map Editing: Towards Efficient Contextual Hallucination Mitigation